Task 2-SemiFrozen Embeddings.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Transfer Learning in NLP : semi-frozen embeddings\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Introduction & Problem Statement\n",
    "\n",
    "The idea behind *\"transfer learning\"* is to take a state of the art model pre-trained on a massive dataset and then *fine-tuning* the model on a new related task where the training data is insufficient to train an analogous model. Transfer learning is gaining more and more traction in the NLP world with models such as [BERT](https://arxiv.org/abs/1810.04805), [Elmo](https://arxiv.org/abs/1802.05365) and [ULMFit](https://arxiv.org/abs/1801.06146) showing the enormous potential. \n",
    "\n",
    "To be fair, transfer learning has been around for quite some time but in a more relaxed setting: the usage of *pretrained word embeddings*.\n",
    "\n",
    "Unfortunately, there is a huge problem - particularly when transferring knowledge from one domain to another - the problem of **vocabulary mismatch**. Very often, the target domain contains words which are quite specific and important, but they are missing from the original domain. It is vital to incorporate these words in the model, but their absence from the pretrained embeddings vocabulary makes it problematic.\n",
    "There are three obvious options for modeling them:\n",
    "1. **Discard** the pretrained embeddings, start from scratch with a random init.\n",
    "2. Set the unknown words to the mean of the pretrained embeddings and **keep them frozen**\n",
    "3. Set the unknown words to the mean of the pretrained embeddings and **unfreeze the entire layer**\n",
    "\n",
    "All of these have their drawbacks:   \n",
    "Option 1. discards any notion of transfer learning.   \n",
    "Option 2. treats all unknown words the same and keeps them frozen, thus failing to model them effectively.        \n",
    "In my experiments, Option 3. would eventually lose all prior knowledge and overfit on the training set.\n",
    "\n",
    "Thus, I propose a compromise: **Semi-Frozen Embeddings**. The semi-frozen embeddings module would split the embeddings into two groups: a frozen group and a trainable group. The *frozen group* will consist of the pretrained word embeddings and will be (duh) frozen - it *will NOT be updated during training*. The *trainable group* will contain the new words that need to be finetuned. They are initialized with the mean of the pretrained embeddings and are *updated during training*.\n",
    "\n",
    "\n",
    "We will explore this idea by experimenting on the [\"Hack the News Challenge\"](https://www.datasciencesociety.net/hack-news-datathon-case-propaganda-detection/) for Propaganda Detection Task 2.\n",
    "We will use the great [fast.ai](https://github.com/fastai/fastai) library and implement the new module in [pytorch](https://pytorch.org/). For pretrained word embeddings we will use [Glove](https://nlp.stanford.edu/projects/glove/)."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "\n",
    "## Data loading and preprocessing\n",
    "\n",
    "We'll apply the following steps:\n",
    "1. load up the data\n",
    "2. tokenize the raw texts into lowercased tokens\n",
    "3. split the dataset into training and validation subsets. \n",
    "4. vectorize via tfidf\n",
    "5. discard empty lines"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import pathlib\n",
    "import pandas as pd\n",
    "import numpy as np\n",
    "import spacy\n",
    "%reload_ext autoreload\n",
    "%autoreload 2"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Read the data\n",
    "Download the dataset from https://s3.us-east-2.amazonaws.com/propaganda-datathon/dataset/datasets-v5.zip"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "data_path = pathlib.Path('/home/martin/data/fakenews/datasets-v5/tasks-2-3/train/')\n",
    "\n",
    "def read_data(data_path, is_test=False):\n",
    "    sentences = []\n",
    "    labels = []\n",
    "    for fn in data_path.glob('*.txt'):\n",
    "        text = fn.read_text()\n",
    "        line_it = text.split('\\n')\n",
    "        if not is_test:\n",
    "            fn_labels = pathlib.Path(fn.as_posix().replace('.txt', '.task2.labels'))\n",
    "            labels_text = fn_labels.read_text()\n",
    "            labels_it = labels_text.split('\\n')\n",
    "        else:\n",
    "            labels_it = ('n/a' for _ in range(len(line_it)))\n",
    "        for line, label in zip(line_it, labels_it):\n",
    "            if label:\n",
    "                sentences.append(line)\n",
    "                labels.append(label.split()[-1])\n",
    "\n",
    "    return pd.DataFrame({'text': sentences, 'labels': labels})\n",
    "data = read_data(data_path)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 2. Tokenize\n",
    "from fastai.text import Vocab, Tokenizer, TextList\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "tl = TextList.from_df(data, cols='text')\n",
    "split_tl = tl.split_by_rand_pct(seed=42)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "db = split_tl.label_from_df('labels')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "db = db.databunch()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "vocab = db.train_ds.vocab"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Load pretrained embeddings\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We will load the pretrained word embeddings, extract the word vectors that are common with our vocabulary and initialize the missing word embeddings as the mean of the original pretrained word embeddings."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "from torch.nn import Embedding\n",
    "from gensim.models.keyedvectors import KeyedVectors"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "wv = KeyedVectors.load_word2vec_format('/data/embeddings/glove.100k.300d.txt')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "# calculate the mean of the pretrained word embeddings\n",
    "mean = wv.vectors.mean(axis=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "missing = []\n",
    "missing_words = []\n",
    "pretrained = []\n",
    "for i, w in enumerate(vocab.itos):\n",
    "    if w in wv:\n",
    "        pretrained.append(wv[w][None])\n",
    "    else:\n",
    "        pretrained.append(mean[None])\n",
    "        missing.append(i)\n",
    "        missing_words.append(w)\n",
    "pretrained = np.vstack(pretrained)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((7200, 300),\n",
       " [0, 1, 2, 3, 4, 5, 6, 7, 8, 188],\n",
       " ['xxunk',\n",
       "  'xxpad',\n",
       "  'xxbos',\n",
       "  'xxeos',\n",
       "  'xxfld',\n",
       "  'xxmaj',\n",
       "  'xxup',\n",
       "  'xxrep',\n",
       "  'xxwrep',\n",
       "  'mccarrick'])"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "pretrained.shape, missing[:10], missing_words[:10]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "314"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(missing_words)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Uh, oh - we have **314 missing word**. We need to incorporate them in the model."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## SemiFrozen embeddings"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Internally, the SemiFrozen embeddings layer defines two nn.Embedding layers. \n",
    "Looking up the frozen ids goes to one embedding, looking up the trainable ids goes to another. The intermediate results are then summed to provide the final output.    \n",
    "A *semi-freeze* method is provided which freezes only the embeddings corresponding to the frozen ids of the layer."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "class SemiFrozenEmbedding(nn.Module):\n",
      "\n",
      "    def __init__(self, num_emb, emb_dim, padding_idx, _weight, frozen_ids = None):\n",
      "        \"\"\"\n",
      "        Acts as a replacement for the nn.Embedding module where some of the vectors for the frozen_ids are kept frozen.\n",
      "        :param num_emb:\n",
      "        :param emb_dim:\n",
      "        :param padding_idx:\n",
      "        :param _weight:\n",
      "        :param frozen_ids:\n",
      "        \"\"\"\n",
      "        frozen_ids = frozen_ids if frozen_ids is not None else []\n",
      "        frozen_id_set = set(frozen_ids)\n",
      "        self.embedding_dim, self.num_emb = emb_dim, num_emb\n",
      "        super(SemiFrozenEmbedding, self).__init__()\n",
      "        self.frozen_ids = frozen_ids\n",
      "        frozen_weight = torch.cat( [torch.zeros(padding_idx+1, emb_dim), _weight[self.frozen_ids]], dim=0)\n",
      "        self.frozen_emb = nn.Embedding(len(self.frozen_ids)+padding_idx+1, emb_dim, padding_idx=padding_idx, _weight=frozen_weight)\n",
      "        self.frozen_emb.weight.requires_grad = False\n",
      "        trainable_ids = [x for x in range(num_emb) if x not in frozen_id_set]\n",
      "        trainable_weight = torch.cat([torch.zeros(padding_idx + 1, emb_dim), _weight[trainable_ids]], dim=0)\n",
      "        self.trainable_emb = nn.Embedding(num_emb - len(self.frozen_ids) + padding_idx + 1, emb_dim, padding_idx=padding_idx, _weight=trainable_weight)\n",
      "        self.trainable_map = torch.zeros(num_emb, dtype=torch.long) + padding_idx\n",
      "        self.frozen_map = torch.zeros(num_emb, dtype=torch.long) + padding_idx\n",
      "        frozen_idx = padding_idx + 1\n",
      "        trainable_idx = padding_idx + 1\n",
      "        for i in range(num_emb):\n",
      "            if i == padding_idx:\n",
      "                trainable_idx += 1\n",
      "                continue\n",
      "            elif i in frozen_id_set:\n",
      "                self.frozen_map[i] = frozen_idx\n",
      "                frozen_idx += 1\n",
      "            else:\n",
      "                self.trainable_map[i] = trainable_idx\n",
      "                trainable_idx += 1\n",
      "\n",
      "    def forward(self, *input):\n",
      "        text_input = input[0]\n",
      "        self.trainable_map = self.trainable_map.to(text_input.device)\n",
      "        self.frozen_map = self.frozen_map.to(text_input.device)\n",
      "        trainable = self.trainable_emb(self.trainable_map[text_input])\n",
      "        frozen = self.frozen_emb(self.frozen_map[text_input])\n",
      "        return trainable + frozen\n",
      "\n",
      "    def semi_freeze(self):\n",
      "        self.frozen_emb.weight.requires_grad = False\n",
      "\n",
      "    def semi_unfreeze(self):\n",
      "        self.frozen_emb.weight.requires_grad = True\n",
      "\n"
     ]
    }
   ],
   "source": [
    "from SemiFrozenV2 import SemiFrozenEmbedding\n",
    "import inspect\n",
    "print(inspect.getsource(SemiFrozenEmbedding))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Model\n",
    "Our architecture will consist of a simple embedding->bidir-LSTM->Linear model with some added dropout."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch.nn as nn\n",
    "import torch\n",
    "from fastai.text import WeightDropout, RNNDropout"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "class SimpleModel(nn.Module):\n",
    "    \n",
    "    def __init__(self, emb, lstm_size):\n",
    "        super().__init__()\n",
    "        self.emb = emb\n",
    "        self.lstm = WeightDropout(nn.LSTM(emb.embedding_dim, lstm_size, num_layers=1, bidirectional=True),weight_p=0.2)\n",
    "        self.linear = nn.Linear(lstm_size, 1)\n",
    "        self.rnn_drop = RNNDropout()\n",
    "\n",
    "    def forward(self, inputs):\n",
    "        inputs = inputs.long().t()\n",
    "        emb = self.emb(inputs)\n",
    "        emb = self.rnn_drop(emb)\n",
    "        enc = self.lstm(emb)[-1][-1][-1]\n",
    "        res = self.linear(enc)\n",
    "        res = res.squeeze(-1)\n",
    "        return res\n",
    "    \n",
    "    def reset(self):\n",
    "        pass\n",
    "    \n",
    "    \n",
    "def simple_split(model):\n",
    "    return [[model.emb], [model.lstm], [model.linear]]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[9, 10, 11, 12, 13, 14, 15, 16, 17, 18]"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "semifrozen_ids = [i for i in range(len(vocab.itos)) if i not in missing]\n",
    "semifrozen_ids[:10]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "SimpleModel(\n",
       "  (emb): SemiFrozenEmbedding(\n",
       "    (frozen_emb): Embedding(6888, 300, padding_idx=1)\n",
       "    (trainable_emb): Embedding(316, 300, padding_idx=1)\n",
       "  )\n",
       "  (lstm): WeightDropout(\n",
       "    (module): LSTM(300, 300, bidirectional=True)\n",
       "  )\n",
       "  (linear): Linear(in_features=300, out_features=1, bias=True)\n",
       "  (rnn_drop): RNNDropout()\n",
       ")"
      ]
     },
     "execution_count": 18,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sfe = SemiFrozenEmbedding(pretrained.shape[0], pretrained.shape[1], frozen_ids=semifrozen_ids,_weight=torch.FloatTensor(pretrained),padding_idx=1)\n",
    "model = SimpleModel(sfe, 300)\n",
    "model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "from fastai.text import DataLoader, RNNLearner, Learner, pad_collate\n",
    "from fastai.metrics import accuracy_thresh, fbeta\n",
    "from torch.nn.modules.loss import BCELoss, BCEWithLogitsLoss\n",
    "\n",
    "\n",
    "def wrapped_fbeta(xs ,ys):\n",
    "    return fbeta(xs[None], ys[None],thresh=0.3)\n",
    "\n",
    "class LossWrapper():\n",
    "\n",
    "    def __init__(self, loss):\n",
    "        self.loss = loss\n",
    "\n",
    "    def __call__(self, outputs, targets):\n",
    "        return self.loss(outputs, targets.float())\n",
    "\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [],
   "source": [
    "def fit(learner):\n",
    "    learner.fit_one_cycle(1, 3e-3)\n",
    "    learner.fit_one_cycle(1, 3e-3)\n",
    "    learner.fit_one_cycle(1, 3e-3)\n",
    "    learner.fit_one_cycle(3, 3e-3)\n",
    "    learner.fit_one_cycle(10, 3e-3)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Random Init"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "In the Random init experiment we will use the standard `nn.Embedding` layer with a **random init** to compute the embeddings and train from scratch."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "SimpleModel(\n",
       "  (emb): Embedding(7200, 300, padding_idx=1)\n",
       "  (lstm): WeightDropout(\n",
       "    (module): LSTM(300, 300, bidirectional=True)\n",
       "  )\n",
       "  (linear): Linear(in_features=300, out_features=1, bias=True)\n",
       "  (rnn_drop): RNNDropout()\n",
       ")"
      ]
     },
     "execution_count": 21,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sfe = Embedding(pretrained.shape[0], pretrained.shape[1],padding_idx=1)\n",
    "model = SimpleModel(sfe, 300)\n",
    "model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [],
   "source": [
    "random_learn = Learner(db, model, \n",
    "                       loss_func=LossWrapper(BCEWithLogitsLoss(pos_weight=torch.FloatTensor([3]).cuda())),\n",
    "                      metrics=[accuracy_thresh, wrapped_fbeta])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "Total time: 00:12 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>1.029042</td>\n",
       "      <td>1.067159</td>\n",
       "      <td>0.275214</td>\n",
       "      <td>0.598381</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 00:11 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>1.049066</td>\n",
       "      <td>1.042152</td>\n",
       "      <td>0.474291</td>\n",
       "      <td>0.601381</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 00:12 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>1.000106</td>\n",
       "      <td>1.011226</td>\n",
       "      <td>0.515491</td>\n",
       "      <td>0.601878</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 00:37 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.978392</td>\n",
       "      <td>0.998872</td>\n",
       "      <td>0.589980</td>\n",
       "      <td>0.603536</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>0.949748</td>\n",
       "      <td>0.965295</td>\n",
       "      <td>0.616678</td>\n",
       "      <td>0.609664</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>0.850477</td>\n",
       "      <td>0.979596</td>\n",
       "      <td>0.671721</td>\n",
       "      <td>0.587239</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 02:05 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.835242</td>\n",
       "      <td>0.984983</td>\n",
       "      <td>0.680620</td>\n",
       "      <td>0.585300</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>0.841039</td>\n",
       "      <td>0.985263</td>\n",
       "      <td>0.701714</td>\n",
       "      <td>0.581334</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>0.907812</td>\n",
       "      <td>0.999830</td>\n",
       "      <td>0.672380</td>\n",
       "      <td>0.602800</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>0.827893</td>\n",
       "      <td>1.034017</td>\n",
       "      <td>0.655570</td>\n",
       "      <td>0.607041</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>0.765401</td>\n",
       "      <td>1.038271</td>\n",
       "      <td>0.698418</td>\n",
       "      <td>0.591445</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>5</td>\n",
       "      <td>0.688598</td>\n",
       "      <td>1.034798</td>\n",
       "      <td>0.713909</td>\n",
       "      <td>0.594546</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>6</td>\n",
       "      <td>0.638168</td>\n",
       "      <td>1.139107</td>\n",
       "      <td>0.702703</td>\n",
       "      <td>0.594078</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>7</td>\n",
       "      <td>0.611393</td>\n",
       "      <td>1.063503</td>\n",
       "      <td>0.696440</td>\n",
       "      <td>0.593016</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>8</td>\n",
       "      <td>0.584150</td>\n",
       "      <td>1.158759</td>\n",
       "      <td>0.714568</td>\n",
       "      <td>0.577033</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>9</td>\n",
       "      <td>0.569277</td>\n",
       "      <td>1.182415</td>\n",
       "      <td>0.715887</td>\n",
       "      <td>0.575252</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "fit(random_learn)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Pretrained - Frozen"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "In the pretrained frozen experiment we use the **pretrained embeddings**, but **leave them frozen**."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "SimpleModel(\n",
       "  (emb): SemiFrozenEmbedding(\n",
       "    (frozen_emb): Embedding(2, 300, padding_idx=1)\n",
       "    (trainable_emb): Embedding(7202, 300, padding_idx=1)\n",
       "  )\n",
       "  (lstm): WeightDropout(\n",
       "    (module): LSTM(300, 300, bidirectional=True)\n",
       "  )\n",
       "  (linear): Linear(in_features=300, out_features=1, bias=True)\n",
       "  (rnn_drop): RNNDropout()\n",
       ")"
      ]
     },
     "execution_count": 25,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sfe = SemiFrozenEmbedding(pretrained.shape[0], pretrained.shape[1],padding_idx=1, _weight=torch.FloatTensor(pretrained))\n",
    "model = SimpleModel(sfe, 300)\n",
    "model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [],
   "source": [
    "frozen_learn = Learner(db, model, \n",
    "                       loss_func=LossWrapper(BCEWithLogitsLoss(pos_weight=torch.FloatTensor([3]).cuda())),\n",
    "                      metrics=[accuracy_thresh, wrapped_fbeta])\n",
    "frozen_learn.model.emb.frozen_emb.weight.requires_grad\n",
    "frozen_learn.model.emb.trainable_emb.weight.requires_grad = False"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "Total time: 00:11 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>1.036949</td>\n",
       "      <td>1.064255</td>\n",
       "      <td>0.280817</td>\n",
       "      <td>0.598381</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 00:11 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>1.030553</td>\n",
       "      <td>0.995903</td>\n",
       "      <td>0.546473</td>\n",
       "      <td>0.598381</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 00:10 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.962933</td>\n",
       "      <td>0.934779</td>\n",
       "      <td>0.680290</td>\n",
       "      <td>0.607184</td>\n",
       "      <td>00:10</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 00:33 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.986335</td>\n",
       "      <td>0.926020</td>\n",
       "      <td>0.700066</td>\n",
       "      <td>0.607717</td>\n",
       "      <td>00:10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>0.916995</td>\n",
       "      <td>0.873688</td>\n",
       "      <td>0.707976</td>\n",
       "      <td>0.637015</td>\n",
       "      <td>00:10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>0.825625</td>\n",
       "      <td>0.860901</td>\n",
       "      <td>0.704680</td>\n",
       "      <td>0.645452</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 01:48 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.833497</td>\n",
       "      <td>0.861387</td>\n",
       "      <td>0.682268</td>\n",
       "      <td>0.645737</td>\n",
       "      <td>00:10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>0.846425</td>\n",
       "      <td>0.858780</td>\n",
       "      <td>0.685564</td>\n",
       "      <td>0.642492</td>\n",
       "      <td>00:10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>0.831957</td>\n",
       "      <td>0.849927</td>\n",
       "      <td>0.732367</td>\n",
       "      <td>0.641886</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>0.807570</td>\n",
       "      <td>0.836830</td>\n",
       "      <td>0.716216</td>\n",
       "      <td>0.643734</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>0.769800</td>\n",
       "      <td>0.821819</td>\n",
       "      <td>0.702703</td>\n",
       "      <td>0.643932</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>5</td>\n",
       "      <td>0.761525</td>\n",
       "      <td>0.829930</td>\n",
       "      <td>0.678642</td>\n",
       "      <td>0.643677</td>\n",
       "      <td>00:10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>6</td>\n",
       "      <td>0.775259</td>\n",
       "      <td>0.822776</td>\n",
       "      <td>0.699077</td>\n",
       "      <td>0.647169</td>\n",
       "      <td>00:10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>7</td>\n",
       "      <td>0.726362</td>\n",
       "      <td>0.837371</td>\n",
       "      <td>0.669084</td>\n",
       "      <td>0.647329</td>\n",
       "      <td>00:10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>8</td>\n",
       "      <td>0.706667</td>\n",
       "      <td>0.832002</td>\n",
       "      <td>0.712591</td>\n",
       "      <td>0.649950</td>\n",
       "      <td>00:10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>9</td>\n",
       "      <td>0.679742</td>\n",
       "      <td>0.834950</td>\n",
       "      <td>0.713579</td>\n",
       "      <td>0.649631</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "fit(frozen_learn)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Pretrained - unfrozen"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "In the Unfrozen experiment we use the **pretrained embeddings**, but **unfreeze the entire layer**."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "SimpleModel(\n",
       "  (emb): SemiFrozenEmbedding(\n",
       "    (frozen_emb): Embedding(2, 300, padding_idx=1)\n",
       "    (trainable_emb): Embedding(7202, 300, padding_idx=1)\n",
       "  )\n",
       "  (lstm): WeightDropout(\n",
       "    (module): LSTM(300, 300, bidirectional=True)\n",
       "  )\n",
       "  (linear): Linear(in_features=300, out_features=1, bias=True)\n",
       "  (rnn_drop): RNNDropout()\n",
       ")"
      ]
     },
     "execution_count": 28,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sfe = SemiFrozenEmbedding(pretrained.shape[0], pretrained.shape[1],padding_idx=1, _weight=torch.FloatTensor(pretrained))\n",
    "model = SimpleModel(sfe, 300)\n",
    "model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [],
   "source": [
    "unfrozen_learn = Learner(db, model, \n",
    "                       loss_func=LossWrapper(BCEWithLogitsLoss(pos_weight=torch.FloatTensor([3]).cuda())),\n",
    "                      metrics=[accuracy_thresh, wrapped_fbeta])\n",
    "unfrozen_learn.model.emb.frozen_emb.weight.requires_grad=True\n",
    "unfrozen_learn.model.emb.trainable_emb.weight.requires_grad=True"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "Total time: 00:13 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>1.018129</td>\n",
       "      <td>1.077604</td>\n",
       "      <td>0.345089</td>\n",
       "      <td>0.598688</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 00:12 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.994179</td>\n",
       "      <td>0.986560</td>\n",
       "      <td>0.572512</td>\n",
       "      <td>0.598381</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 00:12 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.882321</td>\n",
       "      <td>0.908238</td>\n",
       "      <td>0.637113</td>\n",
       "      <td>0.629900</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 00:39 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.844862</td>\n",
       "      <td>0.912007</td>\n",
       "      <td>0.694792</td>\n",
       "      <td>0.610707</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>0.777872</td>\n",
       "      <td>0.876190</td>\n",
       "      <td>0.681279</td>\n",
       "      <td>0.633756</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>0.689861</td>\n",
       "      <td>0.877331</td>\n",
       "      <td>0.738629</td>\n",
       "      <td>0.622419</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 02:06 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.637705</td>\n",
       "      <td>0.901254</td>\n",
       "      <td>0.739618</td>\n",
       "      <td>0.628486</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>0.656464</td>\n",
       "      <td>0.892994</td>\n",
       "      <td>0.714568</td>\n",
       "      <td>0.624104</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>0.600205</td>\n",
       "      <td>0.879418</td>\n",
       "      <td>0.725115</td>\n",
       "      <td>0.627375</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>0.595047</td>\n",
       "      <td>0.905646</td>\n",
       "      <td>0.751483</td>\n",
       "      <td>0.623657</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>0.479090</td>\n",
       "      <td>1.130373</td>\n",
       "      <td>0.735333</td>\n",
       "      <td>0.615082</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>5</td>\n",
       "      <td>0.419607</td>\n",
       "      <td>1.305805</td>\n",
       "      <td>0.741925</td>\n",
       "      <td>0.592370</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>6</td>\n",
       "      <td>0.340979</td>\n",
       "      <td>1.338840</td>\n",
       "      <td>0.742584</td>\n",
       "      <td>0.595536</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>7</td>\n",
       "      <td>0.325194</td>\n",
       "      <td>1.516280</td>\n",
       "      <td>0.737970</td>\n",
       "      <td>0.586759</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>8</td>\n",
       "      <td>0.270852</td>\n",
       "      <td>1.778989</td>\n",
       "      <td>0.747528</td>\n",
       "      <td>0.564778</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>9</td>\n",
       "      <td>0.270078</td>\n",
       "      <td>1.710248</td>\n",
       "      <td>0.746210</td>\n",
       "      <td>0.571311</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "fit(unfrozen_learn)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Pretrained semi-frozen"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Finally, in the semi-frozen experiment we **unfreeze the embeddings for the missing words**, while **leaving the original embeddings frozen**."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "SimpleModel(\n",
       "  (emb): SemiFrozenEmbedding(\n",
       "    (frozen_emb): Embedding(6888, 300, padding_idx=1)\n",
       "    (trainable_emb): Embedding(316, 300, padding_idx=1)\n",
       "  )\n",
       "  (lstm): WeightDropout(\n",
       "    (module): LSTM(300, 300, bidirectional=True)\n",
       "  )\n",
       "  (linear): Linear(in_features=300, out_features=1, bias=True)\n",
       "  (rnn_drop): RNNDropout()\n",
       ")"
      ]
     },
     "execution_count": 33,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sfe = SemiFrozenEmbedding(pretrained.shape[0], pretrained.shape[1],padding_idx=1, \n",
    "                          frozen_ids= semifrozen_ids,\n",
    "                          _weight=torch.FloatTensor(pretrained))\n",
    "model = SimpleModel(sfe, 300)\n",
    "model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [],
   "source": [
    "semifrozen_learn = Learner(db, model, \n",
    "                       loss_func=LossWrapper(BCEWithLogitsLoss(pos_weight=torch.FloatTensor([3]).cuda())),\n",
    "                      metrics=[accuracy_thresh, wrapped_fbeta])\n",
    "semifrozen_learn.model.emb.trainable_emb.weight.requires_grad=True\n",
    "semifrozen_learn.model.emb.semi_freeze()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "Total time: 00:11 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>1.030779</td>\n",
       "      <td>1.052817</td>\n",
       "      <td>0.355636</td>\n",
       "      <td>0.598381</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 00:12 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>1.012239</td>\n",
       "      <td>1.033273</td>\n",
       "      <td>0.435069</td>\n",
       "      <td>0.595643</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 00:11 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.985476</td>\n",
       "      <td>0.967111</td>\n",
       "      <td>0.613711</td>\n",
       "      <td>0.593695</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 00:34 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>1.011494</td>\n",
       "      <td>1.017237</td>\n",
       "      <td>0.384970</td>\n",
       "      <td>0.606509</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>1.008234</td>\n",
       "      <td>0.969931</td>\n",
       "      <td>0.666777</td>\n",
       "      <td>0.598381</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>0.954141</td>\n",
       "      <td>0.973938</td>\n",
       "      <td>0.722479</td>\n",
       "      <td>0.598381</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "text/html": [
       "Total time: 01:57 <p><table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy_thresh</th>\n",
       "      <th>wrapped_fbeta</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.939534</td>\n",
       "      <td>0.992071</td>\n",
       "      <td>0.731707</td>\n",
       "      <td>0.598381</td>\n",
       "      <td>00:10</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>0.890306</td>\n",
       "      <td>0.900971</td>\n",
       "      <td>0.686882</td>\n",
       "      <td>0.622626</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>0.887664</td>\n",
       "      <td>0.871179</td>\n",
       "      <td>0.677653</td>\n",
       "      <td>0.637678</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>0.877104</td>\n",
       "      <td>0.951197</td>\n",
       "      <td>0.519446</td>\n",
       "      <td>0.622470</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>0.845691</td>\n",
       "      <td>0.832453</td>\n",
       "      <td>0.680949</td>\n",
       "      <td>0.643938</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>5</td>\n",
       "      <td>0.825332</td>\n",
       "      <td>0.838189</td>\n",
       "      <td>0.668095</td>\n",
       "      <td>0.644616</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>6</td>\n",
       "      <td>0.764899</td>\n",
       "      <td>0.842968</td>\n",
       "      <td>0.741266</td>\n",
       "      <td>0.652699</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>7</td>\n",
       "      <td>0.727933</td>\n",
       "      <td>0.828497</td>\n",
       "      <td>0.715887</td>\n",
       "      <td>0.649394</td>\n",
       "      <td>00:12</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>8</td>\n",
       "      <td>0.742309</td>\n",
       "      <td>0.825067</td>\n",
       "      <td>0.716546</td>\n",
       "      <td>0.647965</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>9</td>\n",
       "      <td>0.699872</td>\n",
       "      <td>0.824616</td>\n",
       "      <td>0.724127</td>\n",
       "      <td>0.646068</td>\n",
       "      <td>00:11</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "fit(semifrozen_learn)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Plot"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "import seaborn as sns\n",
    "sns.set_style('whitegrid')\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.legend.Legend at 0x7f975d0bad30>"
      ]
     },
     "execution_count": 37,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1152x360 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "fig, ax = plt.subplots(1,1, figsize=(16,5), sharex=True)\n",
    "sns.despine(fig)\n",
    "ax.plot(frozen_learn.recorder.val_losses,label='Frozen', lw=4)\n",
    "ax.plot(semifrozen_learn.recorder.val_losses,label='Semifrozen', lw=4)\n",
    "ax.set_xlabel(\"Epoch\", size=18)\n",
    "ax.set_ylabel(\"Val. Loss\", size=18)\n",
    "ax.set_title(\"SemiFrozen embeddings achieve lower validation loss\", size=22)\n",
    "fig.legend(fontsize=18)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<matplotlib.legend.Legend at 0x7f975d04b3c8>"
      ]
     },
     "execution_count": 38,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 1152x360 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "fig, ax = plt.subplots(1,1, figsize=(16,5), sharex=True)\n",
    "sns.despine(fig)\n",
    "ax.plot(random_learn.recorder.val_losses,label='Random', lw=4)\n",
    "ax.plot(unfrozen_learn.recorder.val_losses,label='Unfrozen', lw=4)\n",
    "ax.set_xlabel(\"Epoch\", size=18)\n",
    "ax.set_ylabel(\"Val. Loss\", size=18)\n",
    "\n",
    "ax.set_title('Random init and Unfrozen embeddings validation loss diverges', size=22)\n",
    "fig.legend(fontsize=18)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Predict"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import fastprogress"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import re\n",
    "predict_path = pathlib.Path('/home/martin/data/fakenews/dev-INPUT/tasks-2-3/dev/')\n",
    "res = []\n",
    "files = list(predict_path.glob('*.txt'))\n",
    "for fn in fastprogress.progress_bar(files):\n",
    "    text = fn.read_text()\n",
    "    line_it = text.split('\\n')\n",
    "    article_id = fn.name.replace('article', '').replace('.txt','')\n",
    "    fn_labels = pathlib.Path(fn.as_posix().replace('.txt', '.task2.labels'))\n",
    "    labels_text = fn_labels.read_text()\n",
    "    labels_it = labels_text.split('\\n')\n",
    "    for i, (line, l) in enumerate(zip(line_it,labels_it)):\n",
    "        if l == \"\":\n",
    "            continue\n",
    "        label = semifrozen_learn.predict(line)\n",
    "        score = torch.sigmoid(label[2])\n",
    "        if torch.sigmoid(label[2]) > 0.4:\n",
    "            label = 'propaganda'\n",
    "        else:\n",
    "            label = 'non-propaganda'\n",
    "        res.append({'text': line, 'article_id': article_id, 'line': i+1, 'label':label, 'score': score})"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "res = pd.DataFrame(res)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "pd.DataFrame(res)[[\"article_id\", \"line\", \"label\"]].to_csv('/tmp/res.tsv',sep='\\t', header=None, index=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!subl /tmp/res.tsv"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 104,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "15170"
      ]
     },
     "execution_count": 104,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(data)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}